Deep Learning for Enhancing Visual Content: Advancing Image Quality and Fidelity
| dc.contributor.advisor | Stavness, Ian | |
| dc.contributor.committeeMember | Eramian, Mark | |
| dc.contributor.committeeMember | Shirtliffe, Steve | |
| dc.contributor.committeeMember | Jin, Lingling | |
| dc.contributor.committeeMember | Horsch, Mike | |
| dc.contributor.committeeMember | Liu, Zheng | |
| dc.contributor.committeeMember | McQuillan, Ian | |
| dc.creator | Aslahishahri, Masi | |
| dc.date.accessioned | 2024-01-25T19:11:21Z | |
| dc.date.available | 2024-01-25T19:11:21Z | |
| dc.date.copyright | 2024 | |
| dc.date.created | 2024-01 | |
| dc.date.issued | 2024-01-25 | |
| dc.date.submitted | January 2024 | |
| dc.date.updated | 2024-01-25T19:11:21Z | |
| dc.description.abstract | In today’s digital world, enhancing the visual content of images has become increasingly vital in various domains to address practical limitations of imaging systems, such as low resolution, noise, and degradation arising from inherent limitations of capture devices. Among these techniques, image super-resolution and image-to-image translation are potent approaches. Image super-resolution aims to enhance the resolution and level of detail in low-resolution images, while image-to-image translation involves transforming images while preserving content and capturing desired attributes. These techniques, often leveraging deep learning algorithms like generative adversarial networks, have revolutionized the field by generating visually appealing and realistic images. In remote sensing, acquiring high-resolution and accurate imagery is paramount for land cover classification, image-based plant phenotyping, and environmental monitoring. Image super-resolution techniques play a pivotal role in this context, as they effectively enhance the quality of low-resolution remote sensing images, providing a more detailed view of the Earth’s surface. These enhanced images can contribute to the precision of analyses and interpretations in applications ranging from agricultural monitoring to urban planning. Similarly, image-to-image translation is pivotal in providing nuanced information by transforming aerial images into more detailed maps or translating images across different spectral bands. This enhancement process affords a finer level of detail and specificity, enriching the interpretation and analysis of remote sensing data. Integrating image super-resolution and image-to-image translation methodologies can contribute to a refined understanding of the Earth’s surface dynamics in remote sensing. The research comprising this dissertation focuses on advancing image visual content enhancement techniques within remote sensing, particularly image super-resolution and image-to-image translation. By harnessing deep learning algorithms and devising and evaluating novel architectures, the dissertation aims to enhance the resolution and quality of remote sensing images, enabling more accurate analysis, particularly in aerial images. Public and specialized crop datasets will be employed to execute and assess the performance of image super-resolution techniques. As part of this effort, a specialized dataset comprising paired low-resolution and high-resolution crop images captured by different sensors has been curated and aligned using an automated image pre-processing pipeline. This dataset is the foundation for training supervised deep learning models designed explicitly for image super-resolution. Additionally, the dissertation explores image-to-image translation techniques to highlight features of interest in remote sensing data, especially in images acquired from different modalities. An aerial dataset comprising crop images captured across multiple spectral bands has been created to enable the transformation of images from the visible light spectrum to invisible spectral bands. This transformation can provide breeders with deeper insights into the growth status of plants, aiding in plant phenotyping and analysis. To further push the boundaries of visual content enhancement, this dissertation introduces a novel transformer-based image super-resolution model that leverages an additional high-resolution reference image during the super-resolution process. The integration of this reference image provides crucial detail information and guidance to enhance the resolution and fidelity of the low-resolution input image. Extensive experimentation has been conducted to verify the efficacy of this proposed technique. The results from these experiments demonstrate the proposed approach’s effectiveness compared to other state-of-the-art models. The findings of this dissertation hold the potential to significantly impact remote sensing applications by improving image quality, enabling better image-based analysis, and enhancing environmental monitoring. The enhanced visual content will provide valuable insights for decision-making processes across various domains. Furthermore, the developed techniques contribute to the broader image visual content enhancement research field, expanding our capabilities to generate visually appealing and informative imagery in diverse applications. The methods described in this dissertation establish the foundation for prospective advancements in image quality enhancement, bearing positive implications for image-based plant phenotyping and related domains. | |
| dc.format.mimetype | application/pdf | |
| dc.identifier.uri | https://hdl.handle.net/10388/15463 | |
| dc.language.iso | en | |
| dc.subject | Deep learning, Image enhancement techniques, Image super-resolution, Image-to-image translation, Plant phenotyping, Remote sensing, Aerial crop images | |
| dc.title | Deep Learning for Enhancing Visual Content: Advancing Image Quality and Fidelity | |
| dc.type | Thesis | |
| dc.type.material | text | |
| thesis.degree.department | Computer Science | |
| thesis.degree.discipline | Computer Science | |
| thesis.degree.grantor | University of Saskatchewan | |
| thesis.degree.level | Doctoral | |
| thesis.degree.name | Doctor of Philosophy (Ph.D.) |